1CEM DETECTING BLOOD LABORATORY ERRORS USING A BAYESIAN NETWORK: AN EVALUATION ON LIVER ENZYME TESTS

Sunday, October 18, 2009
Grand Ballroom, Salons 1 & 2 (Renaissance Hollywood Hotel)
Quang A. Le, Pharm.D.1, Gregory B. Strylewicz, PhD2 and Jason N. Doctor, PhD1, (1)University of Southern California, Los Angeles, CA, (2)University of Washington, Seattle, WA

Purpose: To detect errors in blood laboratory results using a Bayesian network and to compare results with validated method for automatic detection errors, LabRespond, and logistic regression model.

Method: In Experiment 1 and 2 using a sample of 5,800 observations from the National Health and Nutrition Examination Survey dataset, large, medium, and small synthetic errors were randomly generated and introduced to liver enzymes (ALT, AST, and LDH) of the dataset. Experiment 1 examined the additive effects of systematic errors, while Experiment 2 investigated non-systematic errors. The outcome of interest was the correct detection of liver enzymes as “error” or “not error.” With the Bayesian network, the outcome was predicted by exploiting probabilistic relationships among AST, ALT, LDH, and gender. In addition to AST, ALT, LDH, and gender, LabRespond required more information on related analytes (GGT, ALP, and total bilirubin) to achieve optimal prediction. For logistic regression model, parameters were determined by stepwise selection among analytes and their interaction terms with gender that were significant at αlpha ≤ 0.05. We assessed performance by examining the area under the receiver-operating characteristics curves using a 10-fold cross validation methodology.

Result: In Experiment 1, the Bayesian network significantly outperformed both LabRespond and logistic regression model on detecting large (z = 4.03, p < 0.001 and z = 8.08, p < 0.001, respectively), medium (z = 2.29, p = 0.01 and z = 4.99, p < 0.001, respectively), and small (z = 1.87, p = 0.03 and z = 1.58, p = 0.05) systematic errors. In Experiment 2, the Bayesian network performed significantly better than LabRespond and the overall logistic regression model on detecting large (z = 1.78, p = 0.04 and z = 5.15, p < 0.001, respectively) and medium (z = 1.64, p = 0.05 and z = 3.14, p < 0.001, respectively) non-systematic errors. However, the predictive performance for the Bayesian network compared with LabRespond and the overall logistic regression model on detecting small non-systematic error were not statistically significant (p = 0.07 and p = 0.06, respectively).

Conclusion: A Bayesian network detects errors better and with less information than existing automated models, suggesting that Bayesian model can be an effective means for reducing medical costs and increasing patient safety in the laboratory.

Candidate for the Lee B. Lusted Student Prize Competition